Analysis of Claude Code cache bugs causing 10-20x token inflation (github.com)

Recent analysis has revealed significant issues with Claude Code related to client-side bugs, specifically concerning cache inefficiencies that lead to a token inflation of 10-20 times. Out of 11 confirmed bugs, two have been fixed in version 2.1.91, while an additional four preliminary findings have surfaced. Notably, the "Output efficiency" system prompt has been removed, indicating potential changes in how output is managed, which could impact performance. Research suggests that approximately 79% of new sessions still experience a cache miss on the first API call, leading to higher costs and lower efficiency for natural language processing tasks. These findings are critical for the AI/ML community as they highlight structural issues within the caching architecture, which could affect the performance of AI models relying on Claude Code. The ongoing bugs (B3-B11) denote risks for developers using this framework, potentially leading to unexpected costs and inefficiencies. Furthermore, the investigation, supported by community collaboration, underscores the importance of transparency in AI system operations and the need for continuous monitoring and debugging practices to address emerging challenges effectively.